Research Article A Novel Algorithm for Feature …downloads.hindawi.com/journals/acisc/2013/515918.pdfA Novel Algorithm for Feature Level Fusion Using SVM Classifier for Multibiometrics-Based

Hindawi Publishing CorporationApplied Computational Intelligence and Soft ComputingVolume 2013 Article ID 515918 11 pageshttpdxdoiorg1011552013515918

Research ArticleA Novel Algorithm for Feature Level Fusion Using SVMClassifier for Multibiometrics-Based Person Identification

Ujwalla Gawande1 Mukesh Zaveri2 and Avichal Kapur3

1 Department of Computer Technology Yeshwantrao Chavan College of Engineering Nagpur 441110 India2Department of Computer Engineering Sardar Vallabhbhai National Institute of Technology Surat India3 Nagar Yuwak Shikshan Sanstha Nagpur India

Correspondence should be addressed to Ujwalla Gawande ujwallgawandeyahoocoin

Received 1 April 2013 Revised 16 June 2013 Accepted 17 June 2013

Academic Editor Zhang Yi

Copyright copy 2013 Ujwalla Gawande et al This is an open access article distributed under the Creative Commons AttributionLicense which permits unrestricted use distribution and reproduction in any medium provided the original work is properlycited

Recent times witnessed many advancements in the field of biometric and ultimodal biometric fields This is typically observed inthe area of security privacy and forensics Even for the best of unimodal biometric systems it is often not possible to achievea higher recognition rate Multimodal biometric systems overcome various limitations of unimodal biometric systems such asnonuniversality lower false acceptance and higher genuine acceptance rates More reliable recognition performance is achievableas multiple pieces of evidence of the same identity are available The work presented in this paper is focused on multimodalbiometric system using fingerprint and iris Distinct textual features of the iris and fingerprint are extracted using the Haar wavelet-based technique A novel feature level fusion algorithm is developed to combine these unimodal features using the Mahalanobisdistance technique A support-vector-machine-based learning algorithm is used to train the system using the feature extractedThe performance of the proposed algorithms is validated and compared with other algorithms using the CASIA iris database andreal fingerprint database From the simulation results it is evident that our algorithm has higher recognition rate and very less falserejection rate compared to existing approaches

1 Introduction

With the advantage of reliability and stability biometricrecognition has been developing rapidly for security andpersonal identity recognition Protecting resources from anintruder is a crucial problem for the owner The multimodalbiometric system integrates more biometrics to improvesecurity and accuracy and hence is capable of handlingmore efficiently the nonuniversality problem of human traitsIn fact it is very common to use a variety of biologicalcharacteristics for identification In fact it is very commonto use a variety of biological characteristics for identifica-tion because different biological characteristics are know-inglyunknowingly used by people to identify a person

Fusion of multiple biometric traits provides more usefulinformation compared to that obtained using unimodalbiometric trait Use of different feature extraction techniquesfrom each modality possibly covers some features those are

not captured by the first method Supplementary informationon the same identity helps in achieving high performance[1] Prevailing practices in multimodal fusion are broadlycategorized as prematching and postmatching fusion [2]Feature level fusion is prematching activity Fusion at thefeature level includes the incorporation of feature sets relatingto multiple modalities The feature set holds richer infor-mation about the raw biometric data than the match scoreor the final decision Integration at feature level is expectedto offer good recognition results But fusion at this level ishard to accomplish The information obtained from differentmodalities may be so heterogeneous that process of fusionis not so easy The sequential and parallel feature fusionresults in high-dimensional feature vector The lesser thenumber of biometric traits the lesser the time requiredfor future fusion The more number of traits covers moreinformation To achieve an increased recognition rate withreduced processing time is the primary objective of this

2 Applied Computational Intelligence and Soft Computing

work (feature fusion) The time required to fuse the featurevectors from different traits is directly proportional to thenumber of modalities In the light of these facts it is betterto have as less traits as possible but the fusion of the featuresextracted from these traits must provide a higher recognitionrate for having successful usage of multiple biometrics forrecognition system compared to the unimodal-based system

Fingerprint and iris are widely used biometric traits Inthis work these traits are used for feature level fusion Haarwavelet is known for its reduced computational complexityThe Haar wavelet technique is employed for feature extrac-tion in both the cases A new method of feature level fusionis proposed implemented and tested with this work Theviewpoint in this fusion is obtaining relatively better perfor-mance than the unimodal features The proposed algorithmfuses the features of individual modalities accurately andefficiently This is evident from simulation results Ongoingsimilar research in this (multimodal) area has been focusedon postclassification ormatching score-level biometric fusionbecause of its simplicity [3 4] However the large-scale bio-metric identification applications still require performanceimprovements Identification is more computational andtime demanding application than the identity verificationTherefore a more specialized classification-based biometricsystem should be approached in order not only to achievethe desired performance improvement but also to decreasethe execution time [1 2] Commonly used classifiers fordifferent biometrics are support vector machines (SVMs)with different kernels (especially Gaussian and polynomials)Gaussian mixture models-based classifiers neural networksand multilayer perceptron [5ndash8] Most of them providedsignificant performance improvements but their results arestrongly dependent on the available datasets [6] Featurelevel fusion trained by the SVM-based classifier is used inthe proposed work for evaluating the performance of theproposed system The idea behind the SVM is to map theinput vectors into a high-dimensional feature space using theldquokernel trickrdquo and then to construct a linear decision functionin this space so that the dataset becomes separated witha maximum margin [6ndash9] Quite often numbers of featureextraction techniques implemented by multimodal systemsare generally equal to the number of biometrics underconsideration Use of single feature extraction technique forextracting features of both contributing traits makes theframework stronger This research work aims at reducing thefalse rejection false acceptance and training and testing timefor reliable recognition

The remainder of this paper is organized as followsSection 2 briefly presents the review of existing multimodalbiometric systems Section 3 describes the extraction of fea-tures fromfingerprint and iris usingHaarwavelet-based tech-nique their fusion using novel algorithm and classificationusing SVM Section 4 presents results of experimentationSection 5 summarizes the recognition performance of theproposed algorithms with existing recognition and fusionalgorithms Finally Section 6 concludes our research andproposes a scope for future work

2 Literature Review

In recent times multimodal biometric systems have attractedthe attention of researchers A variety of articles can befound which propose different approaches for unimodaland multimodal biometric systems which are [1 2 10ndash14] Multimodal fusion has the synergic effect enhancingthe value of information First multimodal biometric wasproposed by Jain and Ross 2002 [10] Later many scientiststhrive for plenty of research in multimodal biometric systemA great deal of academic research was devoted to it Anumber of published works demonstrate that the fusionprocess is effective because fused scores provide much betterdiscrimination than individual scores [1 2] Voluminousliterature deals with a variety of techniques making thefeatures more informative [9ndash12] Single input and multiplealgorithms for feature extraction [9] or multiple samples andsingle feature extraction algorithm [15] or utilizing two ormore different modalities [16] are commonly discussed inrecent times It was found in [17] that the empirical relationin multimodal biometrics can improve the performance Butthese improvements involve the cost of multiple sensors ormultiple algorithms which ultimately reflects higher instal-lation and operational cost

In most cases multimodal fusion can be categorized intotwo groups prematching fusion and postmatching fusionFusion prior to matching integrates pieces of evidence beforematching It includes sensor level fusion [10] and feature levelfusion [18 19] Fusion after matching integrates pieces ofevidence after matching It includes match score level [20ndash22] rank level [23 24] and decision level [3] Fusion atthe decision level is too rigid since only a limited amountof information is available at this level [25] Integration atthe matching score level is generally preferred due to theease in accessing and combining matching scores Sincethe matching scores generated by different modalities areheterogeneous it is essential to normalize the scores beforecombining them Normalization is computationally expen-sive [17] Choosing inappropriate normalization techniquemay result in low recognition rate For harmonious andeffective working of the system careful selection of differentenvironments different set of traits different set of sen-sors and so forth are necessary [17] The features containricher information about the input biometric data than thematching and decision scores Integration at the featurelevel should provide better recognition results than otherlevels of integration [18 19 25] However feature space andscaling differences make it difficult to homogenize featuresfrom different biometrics Fused feature vector sometimesmay result in increased dimension compared to unimodalfeatures For example fused vector is likely to end up withtwice the dimension of unimodal features For fusion toachieve the claimed performance enhancement fusion rulesmust be chosen based on the type of application biometrictraits and level of fusion Biometric systems that integrateinformation at an early stage of processing are believed to bemore effective than those systems which perform integrationat a later stage

Applied Computational Intelligence and Soft Computing 3

In what follows we introduce a brief review of somerecent researchesThe unimodal iris system unimodal palm-print system and multibiometric system (iris and palmprint)are presented in [20] They worked on the matching scoreon the basis of similarity of the query feature vector withthe template vector Besbes et al [3] proposed a multimodalbiometric system using fingerprint and iris featuresThey usea hybrid approach based on (1) fingerprint minutiae extrac-tion and (2) iris template encoding through a mathematicalrepresentation of the extracted Iris region This approachwas based on two recognition modalities and every partprovides its own decision The final decision was obtainedbyANDing the unimodal decision Experimental verificationof the theoretical claim is missing in their work Aguilaret al [26] worked on multibiometric using a combinationof fast Fourier transform (FFT) and Gabor filters enhancefingerprint imaging Successively a novel stage for recog-nition using local features and statistical parameters wasused They used the fingerprints of both thumbs Each fin-gerprint was separately processed and the unimodal resultswere combined to obtain final fused result Yang and Ma[22] used fingerprint palm print and hand geometry toimplement personal identity verification The three imageswere derived from the same image They implementedmatching score fusion to establish identity performing firstfusion of the fingerprint and palm-print features and later amatching-score fusion between the multimodal system andthe unimodal palm geometry An approach suggested in[27] shows improved data fusion for face fingerprint andiris images The approach was based on the eigenface andthe Gabor wavelet methods incorporating the advantagesof the single algorithm They recommended a new fusionsystem that exhibited improved performance Baig et al [4]worked on the state-of-the-art framework for multimodalbiometric identification system It is adaptable to any kindof biometric system Faster processing was benefits of theirsystem A framework for fusion of the iris and fingerprint wasdeveloped for verification Classification was based on singleHamming distance An authentication method presented byNagesh Kumar et al [21] focuses on multimodal biometricsystem with two features that is face and palm printIntegrated feature vector resulted in the robustness of theperson authentication The final assessment was done byfusion at the matching score level Unimodal scores werefused after matching Maurer and Baker have presented afusion architecture based on Bayesian belief networks forfingerprint and voice [28] The features were modelled usingstatistical distributions

A frequency-based approach resulting in a homogeneousbiometric vector integrating iris and fingerprint data isworked out in [25] Successively a hamming-distance-basedmatching algorithm dealt with the unified homogenous bio-metric vector Basha et al [23] implemented fusion of iris andfingerprint They used adaptive rank level fusion directly atverification stage Ko [29]worked on the fusion of fingerprintface and iris Various possibilities of multimodal biometricfusion and strategies were discussed for improving accuracyEvaluation of image quality from different biometric andtheir influence on identification accuracy was also discussed

Jagadeesan et al [18] prepared a secured cryptographic keyon the basis of iris and fingerprint features Minutiae pointswere extracted from fingerprint Similarly texture propertieswere extracted from iris Feature level fusion was furtheremployed 256-bit cryptographic key was the outcome of thefusion Improvement in authentication and security using256-bit encryption was claimed as a part of their result in[18] A multimodal biometric-based encryption scheme wasproposed by [19] They combine features of the fingerprintand iris with a user-defined secret key

As far as multimodal biometric-based encryptionschemes are concerned there are very few proposals inthe literature Nandakumar and Jain [30] proposed a fuzzyvault-based scheme that combines iris with fingerprints Asexpected the verification performance of themulti-biometricvault was better than the unimodal biometric vaults Butthe increase in entropy of the keys is only from 40 bits (forindividual fingerprint and iris modalities) to 49 bits (forthe multibiometric system) Nagar et al [31] proposed afeature-level fusion framework to simultaneously protectmultiple templates of a user as a single secure sketch Theyimplemented framework using two well-known biometriccryptosystems namely fuzzy vault and fuzzy commitment

In contrast to the approaches found in the literatureand detailed earlier the proposed approach introduces aninnovative idea to fuse homogenize size feature vector offingerprint and iris at the feature level A fusion processimplemented in this work is based on the Mahalanobisdistance This fusion has the advantage of reduction infused feature vector size which is the main issue (of highdimensions) in feature level fusion Our proposed approachof feature fusion outperforms the suggested technique by[3] and related comparisons against the unimodal elementsExtraction of features from multimodalities their fusionusing distinct process and classification using SVM are thecore of this work The proposed fusion method is not usedearlier in any research The novelty of the work is to create asingle template from two biometric modalities and the use ofSVM for recognition purpose

3 Proposed Multimodal Biometric System

The proposed approach implements an innovative idea tofuse the features of two different modalitiesmdashfingerprintand iris The proposed system functions are grouped in thefollowing basic stages

(i) preprocessing stage which obtains region of interest(ROI) for further processing

(ii) feature extraction stage which provides the resultingfeatures using Haar wavelet from ROI

(iii) fusion stage which combined the corresponding fea-ture vectors of unimodalThe features so obtained arefused by innovative technique using the MahalanobisdistanceThese distances are normalized using hyper-bolic and are fused by applying the average sum ruletanh


(a) (b)

Figure 1 Preprocessing of Fingerprint Image (a) Input Fingerprint Image (b) Histogram equalized Image

(iv) the classification stage which operates on fingerprintfeature vector iris feature vector and fused featurevector for an entire subject separately We applied anSVM strategy for classification to get the results ofacceptancerejection or identification decision

These techniques are explained as follows

31 Fingerprint Feature Extraction Fingerprint image ismade available from the image acquisition system Realfingerprint images are rarely of perfect quality They may bedegraded by noise due to many factors including variationsin skin impression condition and noise of capturing devicesQuality of some images is really bad preprocessing improvesthem to the extent required by feature extraction Basi-cally fingerprint recognition is performed by minutiae-basedmethod and image-based method Most of the minutiae-based methods require extensive preprocessing operationssuch as normalization segmentation orientation estimationridge filtering binarization and thinning [32] Minutiaedetection is then performed to reliably extract the minutiaefeatures For extracting the features of fingerprint we used aHaar wavelet-based technique This image-based techniquehas its roots in frequency analysis (wavelet) An importantbenefit of this technique is less computational time therebymaking it more suitable for real-time applications Thistechnique requires less preprocessing and works fine evenwith low-quality images

In this paper the fingerprint image was first preprocessedusing normalization histogram equalization and averagefiltering to improve the image quality Normalization of theinput fingerprint image has prespecified mean and varianceHistogram equalization (HE) technique has the basic idea tomap the gray levels based on the probability distribution ofthe input gray levels HE flattens and stretches the dynamicrange of the imagersquos histogram [11] This results in the overallcontrast improvement of the image as shown in Figure 1 It

produces an enhanced fingerprint image that is useful forfeature extraction

311 Haar Wavelet-Based Technique for Extracting Finger-print Features The wavelet transform is a mathematical toolbased onmany-layer function decomposition After applyingwavelet transform a signal can be described by many waveletcoefficients which represent the characteristics of the signalIf the image has distinct features with some frequency anddirection the corresponding subimages have larger energiesin wavelet transform For this reason wavelet transform hasbeen widely used in signal processing pattern recognitionand texture recognition [33] By applying wavelet transformvital information of original image is transformed into acompressed image without much loss of information Haarwavelet decomposed the fingerprint image by mean anddeviation Wavelet coefficients of the image consist of foursubbands each with a quarter of the original area Theup left subimage composed of the low-frequency parts forboth row and column is called approximate image Theremaining three images containing vertical high frequencies(down left) horizontal high frequencies (up right) and highfrequencies in both directions (down right) are detailedimages [34]

If 119891(119909 119910) represents an image signal its Haar wavelettransform is equal to two 1D filters (119909-direction and 119910-direction) As shown in Figure 2(a) where LL representslow-frequency vectors (approximate) HL represents high-frequency vectors in horizontal direction LH representshigh-frequency vectors in vertical direction and HH repre-sents diagonal high-frequency vectors After first decomposi-tion LL quarter that is approximate component is submittedfor next decomposition In this manner the decomposition iscarried out four times as shown in Figure 2(b) This reducesarray size by 1 times 60 along 119909- as well as 119910-direction Theoriginal image of 160 times 96 is reduced to 10 times 6 after fourthdecomposition From this image a single 1 times 60 feature vector


LL

LH

HL

HH

(a)

LLLH

LHLH

LH

HL HLHL

HLHH

HHHH

HH

(b)

Figure 2 (a) Wavelet decomposition and (b) 4-level wavelet decomposition

(a) Iris after boundaries detected (b) Iris image after noise removal (c) Normalized Image

Figure 3 Iris Localization segmentation and normalization

is extracted by row-wise serialization This itself is treated asextracted feature vector for fingerprint

32 Iris Feature Extraction Theprocess of extracting featuresfrom the iris image is discussed in this section The irisimage must be preprocessed before using it for the featureextraction purpose The unwanted data including eyelidpupil sclera and eyelashes in the image should be excludedTherefore the preprocessing module for iris is required toperform iris segmentation iris normalization and imageenhancement Segmentation is an essentialmodule to removenonuseful information namely the pupil segment and thepart outside the iris (sclera eyelids and skin) It estimates theiris boundary For boundary estimation the iris image is firstfed to the canny algorithm which generates the edge map ofthe iris image The detected edge map is then used to locatethe exact boundary of pupil and iris using Hough transforms[12] The next step is to separate eyelid and eyelashes Thehorizontal segmentation operator and image binarizationwere used to extract the eyelid edge information The eyelidsspan the whole image in the horizontal directionThe averageof vertical gradients is larger in the area with eyelid boundaryThe longest possible horizontal line is demarking line foreyelids This is used as the separator segmenting it intotwo parts The eyelid boundaries were modelled with theparabolic curves according to the determined edge pointsIn the process of normalization the polar image of iris istranslated into the Cartesian frame depicting it as rectangularstrip as shown in Figure 3 It is done using Daugmanrsquos rubbersheet model [13] Finally histogram equalization was used forimage enhancement

321 Haar Wavelet-Based Technique for Extracting Iris Fea-tures The enhanced normalized iris image is then used to

extract identifiable features of iris pattern using three-leveldecomposition by Haar wavelet transform Haar waveletis selected in this work because of its ability of capturingapproximate information along with retention of detailedtexture The normalized image of 240 times 20 is decom-posed first along the row (for each row) and then thecolumn (for each column) It produces four regionsmdashleft-up quarter (approximate component) right-up (horizontalcomponent) bottom-left (vertical component) and bottom-right (diagonal component) The mean and deviations ofthe coefficients are available after decomposition We againdecompose approximate quarter into four subquarters Thisdecomposition is employed three times Four sub images ofsize 30 times 2 are obtained The approximate component after3rd decomposition is representative feature vector The firstrow of 1 times 30 is appended by second row of 1 times 30 for gettingthe feature vector of 1 times 60

33 Fusion of Iris and Fingerprint Feature Vectors Thefeature vectors extracted from encoded input images arefurther combined to the new feature vector by proposinga feature level fusion technique The present work includesan innovative method of fusion at the feature level TheMahalanobis distance technique is at the core of this fusionThis makes it distinct from different methods reported inthe literature The extracted features from fingerprint andiris are homogeneous each vector is of size 1 times 60 elementsThese two homogenous vectors are processed pragmaticallyto produce the fused vector of same order that is 1 times 60elements As most of the feature fusion in the literature isperformed serially or parallel it ultimately results in a high-dimensional vectorThis is the major problem in feature levelfusionThe proposed algorithm generates the same size fusedvector as that of unimodal and hence nullified the problem


Query iris image Query fingerprint image

Reference database of Haar wavelet-based method

Feature extraction

Features by Haar wavelet-based method (F1)



Most similar feature vector using the Mahalanobis distance technique

Distance between query and most similar feature vector in database

Mean of each element in two difference vectors

Fused feature vector

Classification (SVM)

DF1 DF2

Figure 4 Architecture for feature fusion

of high dimensionThe fusion process is depicted in Figure 4and is explained as follows

(1) Features fingerprint and iris of the query images areobtained

(2) The nearest match for query feature vectors of finger-print and iris is selected from 4times 100 reference featurevectors of fingerprints and iris using andMahalanobisdistance

(3) The Mahalanobis distance (119872119889) between a sample119909 and a sample 119910 is calculated using the followingequation

119872119889(119909 119910)2= (119909 minus 119910)

1015840119878minus1 (119909 minus 119910) (1)

where 119878 is the within-group covariancematrix In thispaper we assume a diagonal covariance matrix Thisallows us to calculate the distance using only themeanand the variance The minimum distance vector isconsidered as the most similar vector

(4) The difference between query and its most similarvector is computed for fingerprint and iris which areagain of size 1 times 60

(5) The elements of these difference vectors are normal-ized by Tanh to scale them between 0 and 1 Scalingof the participating (input) features to the same scaleensures their equal contribution to fusion

(6) The mean value is calculated for these two differencevectors for each component This yields a new vectorof 1 times 60 That itself is the fused feature vector

This fused vector is used for training using SVMThismethodsaves training and testing time while retaining the benefits offusion

34 Support Vector Machine Classifier SVM has demon-strated superior results in various classification and patternrecognition problems [35 36] Furthermore for severalpattern classification applications SVM has already beenproven to provide better generalization performance thanconventional techniques especially when the number of inputvariables is large [37 38] With this purpose in mind weevaluated the SVM for our fused feature vector

The standard SVM takes a set of input data It is apredictive algorithm to pinpoint the class to which the inputbelongsThismakes the SVManonprobabilistic binary linearclassifier [39] Given a set of training samples eachmarked asbelonging to its categories an SVM training algorithm buildsa model that assigns new sample to one category or anotherFor this purpose we turn to SVM for validating our approachTo achieve better generalization performance of the SVMoriginal input space is mapped into a high-dimensional dotproduct space called the feature space and in the featurespace the optimal hyperplane is determined as shown inFigure 5 The initial optimal hyperplane algorithm proposedby Vapnik [40] was a linear classifier Yet Boser et al [39]suggested a way to create nonlinear classifiers by applyingthe kernel trick to extend the linear learning machine tohandle nonlinear cases We aimed to maximize the marginof separation between patterns to have a better classificationresultThe function that returns a dot product of twomappedpatterns is called a kernel function Different kernels can beselected to construct the SVM The most commonly usedkernel functions are the polynomial linear and Gaussianradial basis kernel functions (RBF) We employed two typesof kernels for experimentation namely radial basis functionkernel and polynomial kernel

The separating hyperplane (described by 119908) is deter-mined by minimizing the structural risk instead of theempirical error Minimizing the structural risk is equivalent


Margin

A

B

w middot x + b = minus1 w middot x + b = +1

w middot x + b = 0

Figure 5 Linearly separable patterns

to seeking the optimal margin between two classes Theoptimal hyperplane can be written as a combination of afew feature points which are called support vectors of theoptimal hyperplane SVM training can be considered asthe constrained optimization problem which maximizes thewidth of the margin and minimizes the structural risk

min119908119887

1

2119908119879119908 + 119862

119873

sum119894=1

120576119894

subject to 119910119894 (119882119879120593 (119909119894) + 119887) ge 1 minus 120576119894

120576119894 ge 0 forall119894

(2)

where 119887 is the bias 119862 is the trade-off parameter 120576119894 is the slackvariable which measures the deviation of a data point fromthe ideal condition of pattern and 120593(sdot) is the feature vectorin the expanded feature space The penalty parameter ldquo119862rdquocontrols the trade-off between the complexity of the decisionfunction and the number of wrongly classified testing pointsThe correct119862 value cannot be known in advance and awrongchoice of the SVM penalty parameter 119862 can lead to a severeloss in performance Therefore the parametric values areusually estimated from the training data by cross-validationand exponentially growing sequences of 119862

119908 =119873SV

sumSV=1120572SV119910SV120593 (119909SV) (3)

where 119873SV is the number of SVs 119910SV isin minus1 +1 is thetarget value of learning pattern 119909SV and 120572SV is the Lagrangemultiplier value of the SVs Given a set of training sampleseach marked as belonging to its categories an SVM trainingalgorithm builds a model that assigns new sample to one ofthe two available zones

Classification of the test sample 119911 is performed by

119910 = sgn(119873SV

sumSV=1120572SV119910SV119870(119909SV 119911)) (4)

where 119870(119909SV 119911) is the functional kernel that maps theinput into higher-dimensional feature space Computational

complexity and classification time for the SVM classifiersusing nonlinear kernels depend on the number of supportvectors (SVs) required for the SVM

The effectiveness of SVM depends on the kernel usedkernel parameters and a proper soft margin or penalty119862 value [41ndash43] The selection of a kernel function is animportant problem in applications although there is notheory to tell which kernel to use In this work two typesof kernels are used for experimentation namely radial basisfunction kernel and Polynomial kernelThe RBF requiresless parameters to set than a polynomial kernel Howeverconvergence for RBF kernels takes longer than for the otherkernels [44]

RBF kernel has the Gaussian form of

119870(119909119894 119909119895) = 119890

minus[(119909119894 minus 119909119895)

2

]

21205902

(5)

where 120590 gt 0 is a constant that defines the kernel widthThe polynomial kernel function is described as

119870(119909119894 119909119895) = (119909119894 sdot 119909119895 + 1)119889

(6)

where 119889 is a positive integer representing the constant ofpolynomial degree

Before training for SVM radial basis function twoparameters need to be found 119862 and 120574 119862 is the penaltyparameter of the error term and 120574 is a kernel parameter Bycross-validation the best 119862 and 120574 values are found to be 80and 01 respectively Similarly polyorder value needs to befound for 119901 for polynomial SVM After regress parametertuning the best 119901 found was 80 Research has found thatthe RBF kernel is superior to the polynomial kernel for ourproposed feature level fusion method

4 Experimental Results

Evaluation of proposed algorithm is carried out in threedifferent sets of experiments The database is generated for100 genuine and 50 imposter sample cases Four imagesof iris as well as fingerprint are stored for each person inthe reference database Similarly one image of each iris andfingerprint for all 100 cases is stored in query database In thesame manner data for 50 cases in reference and 50 cases inthe query is appended for imposters Fingerprint images arereal and iris images are from CASIA database The mutualindependence assumption of the biometric traits allows usto randomly pair the users from the two datasets First wereport all the experiments based on the unimodal Thenthe remaining experiments are related to the fused patternclassification strategy Two kernels RBFSVM and PolySVMare used for classification in separate experiments Threetypes of feature vectors first from fingerprints second fromiris and third as their fusion are used as inputs in separateexperimental setups Combination of input features withclassifier makes six different experimentsThe objective of allthese experiments is to provide a good basis for comparisonBy using kernels the classification performance or the data


Table 1 Average GAR () FAR () and response time (mean) in seconds for unimodal and multimodal techniques

Algorithm Number ofsupport vectors

Testingset

Identification accuracy(GAR) FAR Training time

(mean in seconds)Testing time

(mean in seconds)RBFKernel

PolynomialKernel

RBFKernel

PolynomialKernel

RBFKernel

PolynomialKernel

RBFKernel

PolynomialKernel

Fingerprint 400 100 87 85 4 8 522 478 02 021Iris 400 100 88 84 2 4 325 494 025 016Fusion 400 100 94 93 0 0 327 398 012 019

separability in the high-dimensional space gets improvedSince this is a pattern recognition approach the database isdivided into two parts training and testing This separationof the database into training and test sets was used for findingthe average performance rating For unimodal biometricrecognition 4 features per user were used in training and oneimage per user is used for testing the performance of thesystem For multimodal four fused features per user wereused for training andone is used for testingThis separation ofthe database into training and test sets was used for findingthe average performance results of genuine acceptance rate(GAR) false acceptance rate (FAR) training and testing time

The results obtained are tabulated in Table 1 The resultsindicate that the average rate of correct classification for fusedvector is found to be 94 in case of RBF kernel for 100classes The GAR of 87 and 88 for unimodal fingerprintand iris respectively is obtained Also the average rate ofcorrect classification for fused vector is found to be 93 incase of polykernel for the same dataset It is found to be 85and 84 for unimodal fingerprint and iris respectively Falseacceptance rate (FAR) reached atoneable minima that is 0using fused feature vector with both kernel classifiersWe canalso observe that the RBF-SVM with fused data is the bestcombination for both GAR and FAR

Moreover the results reported here will help us to betterassess the results produced by unimodal and multimodalin terms of training and testing time Time in secondsrequired for this training and testing for unimodal andmultimodal is also tabulated in Table 1 The best performingmethod that is fused feature required training time of 327 sby RBFSVM and 398 s by PolySVM The minimum timerequired for training is for iris byHaar wavelet-basedmethod(325 s) The results revealed that the mean training timefor RBFSVM and for fused data is slightly greater thanunimodal iris Fused feature vector outperforms unimodalfeatures in both cases for testing It took 012 s using RBFSVMand 019 s using PolySVM Our experimental results provideadequate support for recommending fused feature vectorwith RBFSVM as the best combination for high recognitionIt has highest GAR of 94 and FAR of 0 It has least testingtime of 012 sThis combination is the best performer amongstsix different experiments carried out The GAR of 93 andFAR of 0 with testing time of 019 s for PolySVM is thesecond best performerThe training time for fused data lacksbehind unimodal iris but it should be tolerated as the trainingis to be carried out only once From Table 1 it can be easilyconcluded that feature level performs better compared tounimodal biometric systems

5 Discussion and Comparison

This paper presents a feature level fusion method for a multi-modal biometric system based on fingerprints and irisesTheproposed approach for fingerprint and iris feature extractionfusion and classification by RBFSVM and PolySVM hasbeen tested for unimodal as well as multimodal identificationsystems using the real fingerprint database and CASIA irisdatabase In greater detail the proposed approach performsfingerprint and iris feature extraction using the Haar waveletbased method These codified features are the representationof unique template We compare our approach with Besbeset al[3] Jagadeesan et al [18] and Conti et al [25]

Besbes et al [3] worked on the same modalities (fin-gerprint and iris) but fusion is at the decision level Theyclaimed the improvement in results because of their fusionThe claim is not supported by experimental findings Theprocedure followed by them is explained here Features offingerprint were extracted by minutiae-based method Theiris pattern was encoded by 2D Gabor filter Matching wasdone by hamming distance Each modality is processedseparately to obtain its decision The final decision of thesystem uses the operator ldquoANDrdquo between decision comingfrom the fingerprint recognition step and that coming fromthe iris recognition one This ANDing operation is fusionat decision level Hence nobody can be accepted unlessboth of the results are positive To compare their approachwith our work we fused their features sequentially and trainthese vectors by RBFSVM and PolySVM for two separateexperiments

Similar attemptmade by Jagadeesan et al [18] extracts thefeatures minutiae points and texture properties from the fin-gerprint and iris images respectivelyThen the extracted fea-tures were combined together with their innovative methodto obtain the fused multi-biometric template A 256-bitsecure cryptographic key was generated from the multi-biometric template For experimentation they employed thefingerprint images obtained from publicly available sources(so we used our real fingerprint database) and the Iris imagesfrom CASIA iris database Training and testing aspects werenot covered in their literature In our work we implementedthis system separately and train by SVM for comparison

The feature level fusion proposed by Conti et al [25] wasbased on frequency-based approach Their generated fusedvector was homogeneous biometric vector integrating irisand fingerprint data A hamming-distance-based matchingalgorithm was used for final decision To compare ourapproach we train their fused vector with SVM Results


Table 2 Comparison of the proposed approach with Besbes et al [3] Jagadeesan et al [18] and Conti et al [25] (FAR FRR and responsetime (mean) in seconds)

Author Methods Modalities FAR (RBFSVM)

FRR (RBFSVM)

FAR(PolySVM)

FRR(PolySVM)

Trainingtime (RBFSVM)

Testingtime(RBFSVM)

Trainingtime(PolySVM)

Testingtime(PolySVM)

Besbes et al[3]

Featurelevel fusion

Fingerprint 14 22 16 22 57 035 589 029Iris 12 15 14 18 482 028 52 032

Fusion 4 10 4 10 43 020 498 024

Jagadeesanet al [18]

Featurelevel fusion


Fusion 0 9 0 9 498 020 501 021

Conti et al[25]

Featurelevel fusion


Fusion 0 8 0 9 458 019 469 022

Proposedsystem

Featurelevel fusion


Fusion 0 6 0 7 327 012 398 019

80

85

90

95

100

0ndash25 26ndash50 51ndash75 76ndash100

Reco

gniti

on ra

te

SubjectProposed feature level fusion + RBFSVMProposed feature level fusion + PolySVMFeature level fusion [18] + RBFSVMFeature level fusion [18] + PolySVMFeature level fusion [25] + RBFSVMFeature level fusion [25] + PolySVMFeature level fusion [3] + RBFSVMFeature level fusion [3] + PolySVM

Figure 6 Comparison of recognition rates

obtained for FAR and FRR are now available for comparisonWe tested their technique on our database for 100 genuineand 50 imposter cases The comparison chart of FAR andFRR and response time are tabulated in Table 2 Improvedperformance is exhibited by proposed feature level fusioncompared to the fusion methods used by other researchersFrom Table 2 it can be easily concluded that feature levelperforms better compared to unimodal biometric systems

This section is the comparison of results between our pro-posedmethod and the approach recommended by [3 18 25]

Here the results of feature level fusion are compared Resultsindicate GAR of 80 at the decision level fusion by ANDingof unimodal decision By fusing their features sequentiallyand training by RBFSVM we found the results of GAR =90 FAR = 4 training time of 43 s and testing time of020 s Also we obtained the results of GAR = 90 FAR= 4 training time of 498 s and testing time of 024 s forpolySVMOn the basis of the experimental results it is provedthat the feature level fusion is better than the decision levelfusion for the concept proposed by [3]The results also revealthat the feature level fusion gives better performance thanthe individual traits The fused feature vector of Jagadeesanet al [18] achieves 0 FAR and 9 FRR with both kernel ofSVMThe training time of 498 s and testing time of 020 s arerequired by RBFSVM and training time of 501 s and testingtime of 021 s are required by PolySVM in [18] for fused vectorof one subject Also the results obtained by Conti et al [25]achieved 0FARby both kernels of SVMTheFRR of 8 and9 is obtained by RBFSVM and PolySVM respectively Themean training time for training the fused feature sets of [25]is 458 s and 469 s by RBFSVM and PolySVM respectivelyThe testing time required by fused vector is 019 s by RBFSVMand 022 s by PolySVM of of [25] In our case GAR is 94FAR is 0 training time is 327 s and testing time is 012 sby RBFSVM Also GAR is 93 FAR is 0 training time is398 s and testing time is 019 s by PolySVMThe recognitionrates and error rates comparison of the proposed approachwith [3 18 25] is shown in Figures 6 and 7 Results obtainedfrom our method of fusion are superior to those of Besbes etal [3] Jagadeesan et al [18] and Conti et al [25] The resultsobtained by our method are best competing the results of thereference work for all the four performance indicators Of thetwo classifiers used in our work RBFSVM stands superior toPolySVM Experimental results substantiate the superiorityof feature level fusion and our proposed method over theapproach of [3 18 25]


1210

86420

FARFRR

Methodology used

Proposed feature level fusion + RBFSVM

Proposed feature level fusion + PolySVM

Feature level fusion [18] + RBFSVM

Feature level fusion [18] + PolySVM





Rate

()

Figure 7 Comparison of error rates

6 Conclusion

Content of information is rich with multimodal biometricsIt is the matter of satisfaction that multimodalities are nowin use for few applications covering large population Inthis paper a novel algorithm for feature level fusion andrecognition system using SVM has been proposed Intentionof this work were to evaluate the standing of our proposedmethod of feature level fusion using theMahalanobis distancetechnique These fused features are trained by RBFSVMand PolySVM separately The simulation results show clearlythe advantage of feature level fusion of multiple biometricmodalities over single biometric feature identification Thesuperiority of feature level fusion is concluded on the basis ofexperimental results for FAR FAR and training and testingtime Of the two RBFSVM performs better than PolySVMThe uniqueness of fused template generated also outperformsthe decision level fusion The improvement in performanceof FAR FRR and response time is observed as comparedto existing researches From the experimental results itcan be concluded that the feature level fusion producesbetter recognition than individual modalities The proposedmethod sounds strong enough to enhance the performanceof multimodal biometricThe proposedmethodology has thepotential for real-time implementation and large populationsupport The work can also be extended with other biomet-ric modalities also The performance analysis using noisydatabase may be performed

References

[1] S Soviany and M Jurian ldquoMultimodal biometric securingmethods for informatic systemsrdquo in Proceedings of the 34thInternational Spring Seminar on Electronic Technology pp 12ndash14 Phoenix Ariz USA May 2011

[2] S Soviany C Soviany and M Jurian ldquoA multimodal approachfor biometric authentication with multiple classifiersrdquo in Inter-national Conference on Communication Information and Net-work Security pp 28ndash30 2011

[3] F Besbes H Trichili and B Solaiman ldquoMultimodal biometricsystem based on fingerprint identification and iris recogni-tionrdquo in Proceedings of the 3rd International Conference on

Information and Communication Technologies From Theory toApplications (ICTTA rsquo08) pp 1ndash5 Damascus Syria April 2008

[4] A Baig A Bouridane F Kurugollu and G Qu ldquoFingerprintmdashIris fusion based identification system using a single hammingdistance matcherrdquo International Journal of Bio-Science and Bio-Technology vol 1 no 1 pp 47ndash58 2009

[5] L Zhao Y Song Y Zhu C Zhang and Y Zheng ldquoFacerecognition based on multi-class SVMrdquo in IEEE InternationalConference on Control and Decision pp 5871ndash5873 June 2009

[6] T C Mota and A C G Thome ldquoOne-against-all-basedmulticlass svm strategies applied to vehicle plate characterrecognitionrdquo in IEEE International Joint Conference on NeuralNetworks (IJCNN rsquo09) pp 2153ndash2159 New York NY USA June2009

[7] R Brunelli and D Falavigna ldquoPerson identification usingmultiple cuesrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 17 no 10 pp 955ndash966 1995

[8] S Theodoridis and K Koutroumbas Pattern Recognition Else-vier 4th edition 2009

[9] B Gokberk A A Salah and L Akarun ldquoRank-based decisionfusion for 3D shape-based face recognitionrdquo in Proceedings ofthe 13th IEEE Conference on Signal Processing and Communica-tions Applications pp 364ndash367 Antalya Turkey 2005

[10] A Jain and A Ross ldquoFingerprint mosaickingrdquo in IEEE Inter-national Conference on Acoustics Speech and Signal Processing(ICASSP rsquo02) pp IV4064ndashIV4067 Rochester NY USA May2002

[11] M Sepasian W Balachandran and C Mares ldquoImage enhance-ment for fingerprint minutiae-based algorithms using CLAHEstandard deviation analysis and sliding neighborhoodrdquo in IEEETransactions on World Congress on Engineering and ComputerScience pp 1ndash6 2008

[12] B S Xiaomei Liu Optimizations in iris recognition [PhDthesis] Computer Science and Engineering Notre Dame Indi-anapolis Ind USA 2006

[13] J Daugman ldquoHow iris recognition worksrdquo IEEE Transactionson Circuits and Systems for Video Technology vol 14 no 1 pp21ndash30 2004

[14] L Hong and A Jain ldquoIntegrating faces and fingerprints forpersonal identificationrdquo IEEE Transactions on Pattern Analysisand Machine Intelligence vol 20 no 12 pp 1295ndash1307 1998

[15] J Fierrez-aguilar L Nanni J Ortega-garcia and D MaltonildquoCombining multiple matchers for fingerprint verification a


case studyrdquo in Proceedings of the International Conference onImage Analysis and Processing (ICIAP rsquo05) vol 3617 of LectureNotes in Computer Science pp 1035ndash1042 Springer 2005

[16] A Lumini and L Nanni ldquoWhen fingerprints are combined withIrismdashaCase Study FVC2004 andCASIArdquo International Journalof Network Security vol 4 no 1 pp 27ndash34 2007

[17] L Hong A K Jain and S Pankanti ldquoCan multibiometricsimprove performancerdquo in IEEEWorkshop on Automatic Identi-fication Advanced Technologies pp 59ndash64 New Jersey NJ USA1999

[18] A Jagadeesan TThillaikkarasi and K Duraiswamy ldquoProtectedbio-cryptography key invention from multimodal modalitiesfeature level fusion of fingerprint and Irisrdquo European Journal ofScientific Research vol 49 no 4 pp 484ndash502 2011

[19] I Raglu and P P Deepthi ldquoMultimodal Biometric EncryptionUsing Minutiae and Iris feature maprdquo in Proceedings of IEEEStudentsrsquoInternational Conference on Electrical Electronics andComputer Science pp 94ndash934 Zurich Switzerland 2012

[20] VC Subbarayudu andMVNK Prasad ldquoMultimodal biomet-ric systemrdquo in Proceedings of the 1st International Conference onEmerging Trends in Engineering and Technology (ICETET rsquo08)pp 635ndash640 Nagpur India July 2008

[21] M Nagesh kumar P K Mahesh and M N ShanmukhaSwamy ldquoAn efficient secure multimodal biometric fusion usingpalmprint and face imagerdquo International Journal of ComputerScience vol 2 pp 49ndash53 2009

[22] F Yang and BMa ldquoA newmixed-mode biometrics informationfusion based-on fingerprint hand-geometry and palm-printrdquoin Proceedings of the 4th International Conference on Image andGraphics (ICIG rsquo07) pp 689ndash693 Jinhua China August 2007

[23] A J Basha V Palanisamy and T Purusothaman ldquoFast multi-modal biometric approach using dynamic fingerprint authenti-cation and enhanced iris featuresrdquo in Proceedings of the IEEEInternational Conference on Computational Intelligence andComputing Research (ICCIC rsquo10) pp 1ndash8 Coimbatore IndiaDecember 2010

[24] M M Monwar and M L Gavrilova ldquoMultimodal biometricsystem using rank-level fusion approachrdquo IEEE Transactions onSystems Man and Cybernetics B vol 39 no 4 pp 867ndash8782009

[25] V Conti C Militello F Sorbello and S Vitabile ldquoA frequency-based approach for features fusion in fingerprint and iris multi-modal biometric identification systemsrdquo IEEE Transactions onSystems Man and Cybernetics C vol 40 no 4 pp 384ndash3952010

[26] G Aguilar G Sanchez K Toscano M Nakano and H PerezldquoMultimodal biometric system using fingerprintrdquo in Interna-tional Conference on Intelligent and Advanced Systems (ICIASrsquo07) pp 145ndash150 Kuala Lumpur Malaysia November 2007

[27] L Lin X-F Gu J-P Li L Jie J-X Shi and Y-Y HuangldquoResearch on data fusion of multiple biometric featuresrdquo inInternational Conference on Apperceiving Computing and Intel-ligence Analysis (ICACIA rsquo09) pp 112ndash115 Chengdu ChinaOctober 2009

[28] D E Maurer and J P Baker ldquoFusing multimodal biometricswith quality estimates via a Bayesian belief networkrdquo PatternRecognition vol 41 no 3 pp 821ndash832 2008

[29] T Ko ldquoMultimodal biometric identification for large userpopulation using fingerprint face and IRIS recognitionrdquo inProceedings of the 34th Applied Imagery and Pattern RecognitionWorkshop (AIPR rsquo05) pp 88ndash95 Arlington Va USA 2005

[30] K Nandakumar and A K Jain ldquoMultibiometric templatesecurity using fuzzy vaultrdquo in IEEE 2nd International Conferenceon Biometrics Theory Applications and Systems pp 198ndash205Washington DC USA October 2008

[31] A Nagar K Nandakumar and A K Jain ldquoMultibiometriccryptosystems based on feature-level fusionrdquo IEEE Transactionson Information Forensics and Security vol 7 no 1 pp 255ndash2682012

[32] K Balasubramanian and P Babu ldquoExtracting minutiae fromfingerprint images using image inversion and bi-histogramequalizationrdquo in SPIT-IEEE Colloquium and International Con-ference on Biometrics pp 53ndash56 Washington DC USA 2009

[33] Y-P Huang S-W Luo and E-Y Chen ldquoAn efficient iris recog-nition systemrdquo in Proceedings of the International Conference onMachine Learning and Cybernetics pp 450ndash454 Beijing ChinaNovember 2002

[34] N Tajbakhsh K Misaghian and N M Bandari ldquoA region-based Iris feature extraction method based on 2D-wavelettransformrdquo in Bio-ID-MultiComm 2009 joint COST 2101 and2102 International Conference on Biometric ID Managementand Multimodal Communication vol 5707 of Lecture Notes inComputer Science pp 301ndash307 2009

[35] D Zhang F Song Y Xu and Z Liang Advanced PatternRecognition Technologies with Applications to Biometrics Medi-cal Information Science Reference New York NY USA 2008

[36] C J C Burges ldquoA tutorial on support vector machines forpattern recognitionrdquo Data Mining and Knowledge Discoveryvol 2 no 2 pp 121ndash167 1998

[37] M M Ramon X Nan and C G Christodoulou ldquoBeamforming using support vector machinesrdquo IEEE Antennas andWireless Propagation Letters vol 4 pp 439ndash442 2005

[38] M J Fernandez-Getino Garcıa J L Rojo-Alvarez F Alonso-Atienza and M Martınez-Ramon ldquoSupport vector machinesfor robust channel estimation inOFDMrdquo IEEE Signal ProcessingLetters vol 13 no 7 pp 397ndash400 2006

[39] B E Boser I M Guyon and V N Vapnik ldquoTraining algorithmfor optimal margin classifiersrdquo in Proceedings of the 5th AnnualACMWorkshop on Computational LearningTheory (COLT rsquo92)pp 144ndash152 Morgan Kaufmann San Mateo Calif USA July1992

[40] VN VapnikTheNature of Statistical LearningTheory SpringerNew York NY USA 1999

[41] C-W Hsu and C-J Lin ldquoA simple decomposition method forsupport vector machinesrdquo Machine Learning vol 46 no 1ndash3pp 291ndash314 2002

[42] J Bhatnagar A Kumar and N Saggar ldquoA novel approach toimprove biometric recognition using rank level fusionrdquo in IEEEConference on Computer Vision and Pattern Recognition pp 1ndash6 Hong Kong China June 2007

[43] J Manikandan and B Venkataramani ldquoEvaluation of multiclasssupport vector machine classifiers using optimum threshold-based pruning techniquerdquo IET Signal Processing vol 5 no 5pp 506ndash513 2011

[44] R Noori M A Abdoli A Ameri Ghasrodashti and M JaliliGhazizade ldquoPrediction of municipal solid waste generationwith combination of support vector machine and principalcomponent analysis a case study of mashhadrdquo EnvironmentalProgress and Sustainable Energy vol 28 no 2 pp 249ndash2582009

Submit your manuscripts athttpwwwhindawicom

Computer Games Technology

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014


Distributed Sensor Networks


Advances in

FuzzySystems

Hindawi Publishing Corporationhttpwwwhindawicom

Volume 2014


ReconfigurableComputing

Hindawi Publishing Corporation httpwwwhindawicom Volume 2014


Applied Computational Intelligence and Soft Computing

thinspAdvancesthinspinthinsp

Artificial Intelligence

HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014

Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014


Electrical and Computer Engineering

Journal of

Journal of

Computer Networks and Communications


Hindawi Publishing Corporation

httpwwwhindawicom Volume 2014

Advances in

Multimedia


Biomedical Imaging


ArtificialNeural Systems

Advances in


RoboticsJournal of



Computational Intelligence and Neuroscience

Industrial EngineeringJournal of


Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014


Human-ComputerInteraction

Advances in

Computer EngineeringAdvances in



work (feature fusion) The time required to fuse the featurevectors from different traits is directly proportional to thenumber of modalities In the light of these facts it is betterto have as less traits as possible but the fusion of the featuresextracted from these traits must provide a higher recognitionrate for having successful usage of multiple biometrics forrecognition system compared to the unimodal-based system

Fingerprint and iris are widely used biometric traits Inthis work these traits are used for feature level fusion Haarwavelet is known for its reduced computational complexityThe Haar wavelet technique is employed for feature extrac-tion in both the cases A new method of feature level fusionis proposed implemented and tested with this work Theviewpoint in this fusion is obtaining relatively better perfor-mance than the unimodal features The proposed algorithmfuses the features of individual modalities accurately andefficiently This is evident from simulation results Ongoingsimilar research in this (multimodal) area has been focusedon postclassification ormatching score-level biometric fusionbecause of its simplicity [3 4] However the large-scale bio-metric identification applications still require performanceimprovements Identification is more computational andtime demanding application than the identity verificationTherefore a more specialized classification-based biometricsystem should be approached in order not only to achievethe desired performance improvement but also to decreasethe execution time [1 2] Commonly used classifiers fordifferent biometrics are support vector machines (SVMs)with different kernels (especially Gaussian and polynomials)Gaussian mixture models-based classifiers neural networksand multilayer perceptron [5ndash8] Most of them providedsignificant performance improvements but their results arestrongly dependent on the available datasets [6] Featurelevel fusion trained by the SVM-based classifier is used inthe proposed work for evaluating the performance of theproposed system The idea behind the SVM is to map theinput vectors into a high-dimensional feature space using theldquokernel trickrdquo and then to construct a linear decision functionin this space so that the dataset becomes separated witha maximum margin [6ndash9] Quite often numbers of featureextraction techniques implemented by multimodal systemsare generally equal to the number of biometrics underconsideration Use of single feature extraction technique forextracting features of both contributing traits makes theframework stronger This research work aims at reducing thefalse rejection false acceptance and training and testing timefor reliable recognition

The remainder of this paper is organized as followsSection 2 briefly presents the review of existing multimodalbiometric systems Section 3 describes the extraction of fea-tures fromfingerprint and iris usingHaarwavelet-based tech-nique their fusion using novel algorithm and classificationusing SVM Section 4 presents results of experimentationSection 5 summarizes the recognition performance of theproposed algorithms with existing recognition and fusionalgorithms Finally Section 6 concludes our research andproposes a scope for future work

2 Literature Review

In recent times multimodal biometric systems have attractedthe attention of researchers A variety of articles can befound which propose different approaches for unimodaland multimodal biometric systems which are [1 2 10ndash14] Multimodal fusion has the synergic effect enhancingthe value of information First multimodal biometric wasproposed by Jain and Ross 2002 [10] Later many scientiststhrive for plenty of research in multimodal biometric systemA great deal of academic research was devoted to it Anumber of published works demonstrate that the fusionprocess is effective because fused scores provide much betterdiscrimination than individual scores [1 2] Voluminousliterature deals with a variety of techniques making thefeatures more informative [9ndash12] Single input and multiplealgorithms for feature extraction [9] or multiple samples andsingle feature extraction algorithm [15] or utilizing two ormore different modalities [16] are commonly discussed inrecent times It was found in [17] that the empirical relationin multimodal biometrics can improve the performance Butthese improvements involve the cost of multiple sensors ormultiple algorithms which ultimately reflects higher instal-lation and operational cost

In most cases multimodal fusion can be categorized intotwo groups prematching fusion and postmatching fusionFusion prior to matching integrates pieces of evidence beforematching It includes sensor level fusion [10] and feature levelfusion [18 19] Fusion after matching integrates pieces ofevidence after matching It includes match score level [20ndash22] rank level [23 24] and decision level [3] Fusion atthe decision level is too rigid since only a limited amountof information is available at this level [25] Integration atthe matching score level is generally preferred due to theease in accessing and combining matching scores Sincethe matching scores generated by different modalities areheterogeneous it is essential to normalize the scores beforecombining them Normalization is computationally expen-sive [17] Choosing inappropriate normalization techniquemay result in low recognition rate For harmonious andeffective working of the system careful selection of differentenvironments different set of traits different set of sen-sors and so forth are necessary [17] The features containricher information about the input biometric data than thematching and decision scores Integration at the featurelevel should provide better recognition results than otherlevels of integration [18 19 25] However feature space andscaling differences make it difficult to homogenize featuresfrom different biometrics Fused feature vector sometimesmay result in increased dimension compared to unimodalfeatures For example fused vector is likely to end up withtwice the dimension of unimodal features For fusion toachieve the claimed performance enhancement fusion rulesmust be chosen based on the type of application biometrictraits and level of fusion Biometric systems that integrateinformation at an early stage of processing are believed to bemore effective than those systems which perform integrationat a later stage













(a) (b)










LL

LH

HL

HH

(a)

LLLH

LHLH

LH

HL HLHL

HLHH

HHHH

HH

(b)












Feature extraction









DF1 DF2






119872119889(119909 119910)2= (119909 minus 119910)

1015840119878minus1 (119909 minus 119910) (1)










Margin

A

B


w middot x + b = 0



min119908119887

1

2119908119879119908 + 119862

119873

sum119894=1

120576119894


120576119894 ge 0 forall119894

(2)


119908 =119873SV

sumSV=1120572SV119910SV120593 (119909SV) (3)



119910 = sgn(119873SV

sumSV=1120572SV119910SV119870(119909SV 119911)) (4)





119870(119909119894 119909119895) = 119890

minus[(119909119894 minus 119909119895)

2

]

21205902

(5)


119870(119909119894 119909119895) = (119909119894 sdot 119909119895 + 1)119889

(6)








Testingset




PolynomialKernel

RBFKernel

PolynomialKernel

RBFKernel

PolynomialKernel

RBFKernel

PolynomialKernel













FRR (RBFSVM)

FAR(PolySVM)

FRR(PolySVM)


Testingtime(RBFSVM)



Besbes et al[3]

Featurelevel fusion


Fusion 4 10 4 10 43 020 498 024


Featurelevel fusion


Fusion 0 9 0 9 498 020 501 021

Conti et al[25]

Featurelevel fusion


Fusion 0 8 0 9 458 019 469 022

Proposedsystem

Featurelevel fusion


Fusion 0 6 0 7 327 012 398 019

80

85

90

95

100


Reco

gniti

on ra

te







1210

86420

FARFRR

Methodology used









Rate

()


6 Conclusion


References























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in















(a) (b)










LL

LH

HL

HH

(a)

LLLH

LHLH

LH

HL HLHL

HLHH

HHHH

HH

(b)












Feature extraction









DF1 DF2






119872119889(119909 119910)2= (119909 minus 119910)

1015840119878minus1 (119909 minus 119910) (1)










Margin

A

B


w middot x + b = 0



min119908119887

1

2119908119879119908 + 119862

119873

sum119894=1

120576119894


120576119894 ge 0 forall119894

(2)


119908 =119873SV

sumSV=1120572SV119910SV120593 (119909SV) (3)



119910 = sgn(119873SV

sumSV=1120572SV119910SV119870(119909SV 119911)) (4)





119870(119909119894 119909119895) = 119890

minus[(119909119894 minus 119909119895)

2

]

21205902

(5)


119870(119909119894 119909119895) = (119909119894 sdot 119909119895 + 1)119889

(6)








Testingset




PolynomialKernel

RBFKernel

PolynomialKernel

RBFKernel

PolynomialKernel

RBFKernel

PolynomialKernel













FRR (RBFSVM)

FAR(PolySVM)

FRR(PolySVM)


Testingtime(RBFSVM)



Besbes et al[3]

Featurelevel fusion


Fusion 4 10 4 10 43 020 498 024


Featurelevel fusion


Fusion 0 9 0 9 498 020 501 021

Conti et al[25]

Featurelevel fusion


Fusion 0 8 0 9 458 019 469 022

Proposedsystem

Featurelevel fusion


Fusion 0 6 0 7 327 012 398 019

80

85

90

95

100


Reco

gniti

on ra

te







1210

86420

FARFRR

Methodology used









Rate

()


6 Conclusion


References























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




(a) (b)










LL

LH

HL

HH

(a)

LLLH

LHLH

LH

HL HLHL

HLHH

HHHH

HH

(b)












Feature extraction









DF1 DF2






119872119889(119909 119910)2= (119909 minus 119910)

1015840119878minus1 (119909 minus 119910) (1)










Margin

A

B


w middot x + b = 0



min119908119887

1

2119908119879119908 + 119862

119873

sum119894=1

120576119894


120576119894 ge 0 forall119894

(2)


119908 =119873SV

sumSV=1120572SV119910SV120593 (119909SV) (3)



119910 = sgn(119873SV

sumSV=1120572SV119910SV119870(119909SV 119911)) (4)





119870(119909119894 119909119895) = 119890

minus[(119909119894 minus 119909119895)

2

]

21205902

(5)


119870(119909119894 119909119895) = (119909119894 sdot 119909119895 + 1)119889

(6)








Testingset




PolynomialKernel

RBFKernel

PolynomialKernel

RBFKernel

PolynomialKernel

RBFKernel

PolynomialKernel













FRR (RBFSVM)

FAR(PolySVM)

FRR(PolySVM)


Testingtime(RBFSVM)



Besbes et al[3]

Featurelevel fusion


Fusion 4 10 4 10 43 020 498 024


Featurelevel fusion


Fusion 0 9 0 9 498 020 501 021

Conti et al[25]

Featurelevel fusion


Fusion 0 8 0 9 458 019 469 022

Proposedsystem

Featurelevel fusion


Fusion 0 6 0 7 327 012 398 019

80

85

90

95

100


Reco

gniti

on ra

te







1210

86420

FARFRR

Methodology used









Rate

()


6 Conclusion


References























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




LL

LH

HL

HH

(a)

LLLH

LHLH

LH

HL HLHL

HLHH

HHHH

HH

(b)












Feature extraction









DF1 DF2






119872119889(119909 119910)2= (119909 minus 119910)

1015840119878minus1 (119909 minus 119910) (1)










Margin

A

B


w middot x + b = 0



min119908119887

1

2119908119879119908 + 119862

119873

sum119894=1

120576119894


120576119894 ge 0 forall119894

(2)


119908 =119873SV

sumSV=1120572SV119910SV120593 (119909SV) (3)



119910 = sgn(119873SV

sumSV=1120572SV119910SV119870(119909SV 119911)) (4)





119870(119909119894 119909119895) = 119890

minus[(119909119894 minus 119909119895)

2

]

21205902

(5)


119870(119909119894 119909119895) = (119909119894 sdot 119909119895 + 1)119889

(6)








Testingset




PolynomialKernel

RBFKernel

PolynomialKernel

RBFKernel

PolynomialKernel

RBFKernel

PolynomialKernel













FRR (RBFSVM)

FAR(PolySVM)

FRR(PolySVM)


Testingtime(RBFSVM)



Besbes et al[3]

Featurelevel fusion


Fusion 4 10 4 10 43 020 498 024


Featurelevel fusion


Fusion 0 9 0 9 498 020 501 021

Conti et al[25]

Featurelevel fusion


Fusion 0 8 0 9 458 019 469 022

Proposedsystem

Featurelevel fusion


Fusion 0 6 0 7 327 012 398 019

80

85

90

95

100


Reco

gniti

on ra

te







1210

86420

FARFRR

Methodology used









Rate

()


6 Conclusion


References























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in






Feature extraction









DF1 DF2






119872119889(119909 119910)2= (119909 minus 119910)

1015840119878minus1 (119909 minus 119910) (1)










Margin

A

B


w middot x + b = 0



min119908119887

1

2119908119879119908 + 119862

119873

sum119894=1

120576119894


120576119894 ge 0 forall119894

(2)


119908 =119873SV

sumSV=1120572SV119910SV120593 (119909SV) (3)



119910 = sgn(119873SV

sumSV=1120572SV119910SV119870(119909SV 119911)) (4)





119870(119909119894 119909119895) = 119890

minus[(119909119894 minus 119909119895)

2

]

21205902

(5)


119870(119909119894 119909119895) = (119909119894 sdot 119909119895 + 1)119889

(6)








Testingset




PolynomialKernel

RBFKernel

PolynomialKernel

RBFKernel

PolynomialKernel

RBFKernel

PolynomialKernel













FRR (RBFSVM)

FAR(PolySVM)

FRR(PolySVM)


Testingtime(RBFSVM)



Besbes et al[3]

Featurelevel fusion


Fusion 4 10 4 10 43 020 498 024


Featurelevel fusion


Fusion 0 9 0 9 498 020 501 021

Conti et al[25]

Featurelevel fusion


Fusion 0 8 0 9 458 019 469 022

Proposedsystem

Featurelevel fusion


Fusion 0 6 0 7 327 012 398 019

80

85

90

95

100


Reco

gniti

on ra

te







1210

86420

FARFRR

Methodology used









Rate

()


6 Conclusion


References























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




Margin

A

B


w middot x + b = 0



min119908119887

1

2119908119879119908 + 119862

119873

sum119894=1

120576119894


120576119894 ge 0 forall119894

(2)


119908 =119873SV

sumSV=1120572SV119910SV120593 (119909SV) (3)



119910 = sgn(119873SV

sumSV=1120572SV119910SV119870(119909SV 119911)) (4)





119870(119909119894 119909119895) = 119890

minus[(119909119894 minus 119909119895)

2

]

21205902

(5)


119870(119909119894 119909119895) = (119909119894 sdot 119909119895 + 1)119889

(6)








Testingset




PolynomialKernel

RBFKernel

PolynomialKernel

RBFKernel

PolynomialKernel

RBFKernel

PolynomialKernel













FRR (RBFSVM)

FAR(PolySVM)

FRR(PolySVM)


Testingtime(RBFSVM)



Besbes et al[3]

Featurelevel fusion


Fusion 4 10 4 10 43 020 498 024


Featurelevel fusion


Fusion 0 9 0 9 498 020 501 021

Conti et al[25]

Featurelevel fusion


Fusion 0 8 0 9 458 019 469 022

Proposedsystem

Featurelevel fusion


Fusion 0 6 0 7 327 012 398 019

80

85

90

95

100


Reco

gniti

on ra

te







1210

86420

FARFRR

Methodology used









Rate

()


6 Conclusion


References























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in






Testingset




PolynomialKernel

RBFKernel

PolynomialKernel

RBFKernel

PolynomialKernel

RBFKernel

PolynomialKernel













FRR (RBFSVM)

FAR(PolySVM)

FRR(PolySVM)


Testingtime(RBFSVM)



Besbes et al[3]

Featurelevel fusion


Fusion 4 10 4 10 43 020 498 024


Featurelevel fusion


Fusion 0 9 0 9 498 020 501 021

Conti et al[25]

Featurelevel fusion


Fusion 0 8 0 9 458 019 469 022

Proposedsystem

Featurelevel fusion


Fusion 0 6 0 7 327 012 398 019

80

85

90

95

100


Reco

gniti

on ra

te







1210

86420

FARFRR

Methodology used









Rate

()


6 Conclusion


References























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in






FRR (RBFSVM)

FAR(PolySVM)

FRR(PolySVM)


Testingtime(RBFSVM)



Besbes et al[3]

Featurelevel fusion


Fusion 4 10 4 10 43 020 498 024


Featurelevel fusion


Fusion 0 9 0 9 498 020 501 021

Conti et al[25]

Featurelevel fusion


Fusion 0 8 0 9 458 019 469 022

Proposedsystem

Featurelevel fusion


Fusion 0 6 0 7 327 012 398 019

80

85

90

95

100


Reco

gniti

on ra

te







1210

86420

FARFRR

Methodology used









Rate

()


6 Conclusion


References























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




1210

86420

FARFRR

Methodology used









Rate

()


6 Conclusion


References























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in









































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in










Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in



Documents

Research Article A Novel Algorithm for Feature …downloads.hindawi.com/journals/acisc/2013/515918.pdfA Novel Algorithm for Feature Level Fusion Using SVM Classifier for Multibiometrics-Based